October 2026 review
Are AI Detectors Accurate? 6 Detectors Tested
6 detectors tested, October 2026
| Detector | Version tested | Mean bypass | Failed strict bypass | Source text | Likely AI | Mixed | Same-panel mean | Change | Page |
|---|---|---|---|---|---|---|---|---|---|
| Originality.ai 5.0 | model-5.0 | 4.0% | 96.0% | 0% | 472 of 550 | 63 of 550 | — | — | /detectors/originality-ai-5/ |
| Pangram | 4.1 | 17.8% | 82.2% | 0% | — | — | 19.2% | -13.1% | /detectors/pangram/ |
| GPTZero | 2026-10-07-base | 46.7% | 53.3% | 0% | — | — | 47.0% | -28.6% | /detectors/gptzero/ |
| Originality.ai Turbo | turbo | 76.7% | 23.3% | 0% | — | — | 77.2% | +1.0% | /detectors/originality-ai/ |
| ZeroGPT | default | 82.9% | 17.1% | 2% | — | — | 83.0% | -11.8% | /detectors/zerogpt/ |
| Winston AI | 5.0 | 88.7% | 11.3% | 0% | — | — | 88.6% | -3.0% | /detectors/winston-ai/ |
This is not a false-positive rate. There is no human-written control set, and these numbers are not comparable to vendor accuracy claims.
Source text is the unhumanized AI sample, not a human-written control, and is excluded from the pooled rate.
50 texts · October 2026 · Methodology · scores.csv.
Mean strict bypass by detector
- Originality.ai 5.0
- Pangram
- GPTZero
- Originality.ai Turbo
- ZeroGPT
- Winston AI
Originality.ai 5.0
4.0%
Mean strict bypass 4.0%, rank 1 of 6.
Pangram
17.8%
Mean strict bypass 17.8%, rank 2 of 6.
GPTZero
46.7%
Mean strict bypass 46.7%, rank 3 of 6.
Originality.ai Turbo
76.7%
Mean strict bypass 76.7%, rank 4 of 6.
ZeroGPT
82.9%
Mean strict bypass 82.9%, rank 5 of 6.
Winston AI
88.7%
Mean strict bypass 88.7%, rank 6 of 6.
Turnitin and Copyleaks are not tested. Turnitin has no public API at this volume.
Methodology
How texts are chosen, what counts as a strict bypass, and how every number is recomputed.
Originality.ai 5.0
Hardest for humanizers to pass in October 2026, at 4.0% mean strict bypass.
StealthGPT
Leader on the October 2026 humanizer leaderboard, at 80.3% mean bypass.
FAQ
Can AI detectors catch humanized text?
HumanizerEval reports mean strict-bypass rates on humanizer output, not a vendor accuracy score. This is not a false-positive rate. There is no human-written control set, and these numbers are not comparable to vendor accuracy claims. The numbers on this page are from the October 2026 review, plus the share of those outputs that failed a strict bypass.
Which detector was hardest for humanizers to pass?
Originality.ai 5.0 had the lowest mean strict-bypass rate, at 4.0%. That is not a claim that it is the most accurate detector. See the Originality.ai 5.0 page.
Does a failed strict bypass mean the text was flagged as AI?
A failed strict bypass is any output that did not receive a human score above the 50% bypass threshold. Mixed is a vendor band that can still clear the 50% human-score threshold; it is not the same as a vendor label of AI-written.
Why are Turnitin and Copyleaks missing?
Turnitin and Copyleaks are not tested. Turnitin has no public API at this volume.